search

CXH-Research / DocShadow-SD7K

[ICCV 2023] A large-scale high-resolution dataset satisfies all important data features about document shadow, covers a large number of document shadow images.
https://cxh-research.github.io/DocShadow-SD7K/
MIT License
202 stars 13 forks source link

Detail about the DFE #9

Closed igodogi closed 10 months ago

igodogi commented 10 months ago

Why you set DFE default config as this? Why decoders as the same? https://github.com/CXH-Research/DocShadow-SD7K/blob/78dd930b21e02ea527ac70651a10ff0055e82795/models/backbone.py#L116C60-L116C60

Is there a better way to expand the depth of DFE?

zinuoli commented 10 months ago

Hi, good day. We tried the number of blocks opposite to the encoder, which is 8, 4, 2, 2. At the same time, we also attempted other combinations. However, these combinations did not enhance the performance (only very slight improvement, or even negative improvement), but the amount of parameters increased significantly. To maintain the ability of inference under high resolution and faster speed, we chose the current config.

Moreover, you may try increasing the number of decoders by yourself, but at present, the number of decoders is the optimal choice. You can consider DFE simply as a U-Net.